Automatic Generation of 3-D Curved Object Models by Using the Learning Mechanism of Neural Networks
The paper addresses the automatic model generation problem that faces us when we attempt to construct a flexible vision system. In the paper, we present an object-modeling method which describes patches composing model object surfaces and uses the described patches as primitive features for the model description. In the model, the patches are recorded in a list structure and their order in the list can be changed according to their different importance in recognizing different objects. A new algorithm for quantifying the importance on the basis of the learning mechanism of neural networks is proposed. It is shown that the model is particularly suitable for interpreting images containing partially occluded objects.
KeywordsTraining Image Model Object Feature Ranking Input Cell Model Patch
Unable to display preview. Download preview PDF.
- T.J. Fan, G. Medioni, and R. Nevatia, “Matching 3-D Objects Using Surface Descriptions”Proc. Int. Conf. Robotics and Automationpp.1400–1406, 1988Google Scholar
- T. Nagata, and H. B. Zha, “Recognizing and Locating a Known Object from Multiple Images”, to appear inIEEE Trans. Robotics and Automation Google Scholar
- Y. H. Pao, Adaptive Pattern Recognition and Neural Networks, Addison-Wesley Publishing Company, Inc., 1989Google Scholar
- T. Skordas, and R. Horaud“Planning a Strategy for Recognizing Partially Occluded Parts”, in Proc. 8th Int. Joint Conf. Pattern Recognitionpp.1080–1083, 1986Google Scholar
- H.B. Zha, T. Nagata, and K. Kumamaru, “Determination of Feature Points of 3-D Curved Surfaces by Computing Surface Saliency from Range Images”, to appear in Proc. IECON’91, 1991Google Scholar